ServiceNow AI Ticket Resolution Assistant
A Chrome extension that turns support tickets into actionable troubleshooting steps, with knowledge-base citations and explicit uncertainty labels.
- Python
- FastAPI
- RAG
- Chrome Extension
Help inside the support workflow
Support consultants need to connect a ticket's symptoms with the right troubleshooting documentation. I built a Chrome side panel that reads the current incident and presents suggested next steps alongside the ticket.
From ticket to cited recommendation
The pipeline searches 1,394 cleaned knowledge-base articles using BGE embeddings and category-aware ranking. DeepSeek turns the retrieved material into structured steps with source citations and SUPPORTED, INFERRED, or UNCERTAIN labels. At this scale, an in-memory NumPy index is enough — brute-force cosine similarity across all 1,394 articles finishes in single-digit milliseconds, so a dedicated vector database would only add operational complexity without a real latency win.
Handling weak evidence
Each recommended step carries its own confidence label — SUPPORTED means it's grounded directly in a cited KB article, INFERRED means the model reasoned beyond what's explicitly written. When retrieval turns up nothing usable at all, the system skips generation entirely and returns an insufficient-evidence response instead of guessing.
Evaluating retrieval quality
On 200 hand-written evaluation tickets, an expected article appeared in the top five results for 189 of them — a 94.5% hit rate, with a mean reciprocal rank of 0.845. These numbers measure how well the system finds the right article, not whether a ticket actually got resolved.